The AI-Driven Workflow: A Necessary Evolution of the SDLC

The software development life cycle, SDLC, has traditionally followed a flow of requirements: Design, Implementation, Testing, and Evaluation in a perpetual circle.

Working within the SDLC, teams have traditionally followed 1 of 3 main workflows:

1. Waterfall - all specifications are gathered at the beginning and everything is delivered at the end.
2. Scrum/Agile - Shorter SDLC cycles, known as sprints, that are typically around 2 weeks.
3. Kanban - More for support or work after initial delivery, Kanban works through tickets in a prioritized order.

These approaches are valuable across various project types because they prioritize keeping people and scope aligned rather than focusing on code generation; consequently, product owners, developers, and architects often dedicate significant time to rituals. While integrating AI speeds up software writing, its dependency on human direction slows the tool itself, resulting in a velocity improvement of only 10%.

Enter the AI Development Life Cycle, (AI DLC). AI DLC focuses on empowering the AI to go faster while maintaining human oversight and encouraging collaborative development throughout the workflow.

AI DLC is broken into 3 phases, instead of the traditional 5: Inception, Construction, and Operation.

• Inception: The team works together to flush out business requirements, prompting the AI with what they want to build. At this stage, the AI will guide the team through requirements gathering, asking questions as it progresses to get clarity from the team without making assumptions.
• Construction: The AI will create architecture diagrams, data models and an execution plan. Similarly, it asks questions at every step to get clarity to finally create code. Since it has all the detailed planning beforehand, the code is generated faster and more accurately.
• Operation: The AI uses historical context to manage Infrastructure as Code, IaC, and deployments.

Note: It's important to follow the AI and validate everything it does. In addition to questions, it will also pause and ask you to review what it created before proceeding. You can ask for changes at any point in the process.

The AI DLC process is simple, but it doesn’t fit with the Waterfall, Scrum, or Kanban workflows. Each one of those has rituals that can hold back the AI’s velocity. So, to alleviate this issue, there is a new workflow that you can follow.

In the Inception phase, you’ll split the work into individual units of work. After you transition to Construction, the team works in “Bolts”, which are measured in days or hours, instead of the single Sprint’s weeks. Lastly, each bolt is tested and packaged to be deployed with the rest of the unit.

Following the AI DLC workflow, most teams experience a 40%-60% improvement in velocity.

Metal Toad and AI DLC

Metal Toad has traditionally followed Scrum/Agile, and one of the core tenets of this approach is “people over process”. So we’ve brought this philosophy to AI DLC.

As an agency, Metal Toad works with many different clients coming from different stages of software development maturity. Agile went along well with this, allowing them to lean in when they liked, while providing fixed check points at the end of sprint demos. This way clients could minimize the time they needed to meet with Metal Toad to 1-2 meetings a sprint.

With AI DLC’s Units and Bolts, that would require more check-ins with clients, which they may be unable to do constantly. To solve this, Metal Toad uses a hybrid of Scrum and AI DLC to facilitate customer involvement/communication and team velocity.

These changes are important.

1. Units of work are decided on during Sprint Planning/Backlog Grooming. At this point, we fill out initial prompts that are fed to the AI to guide it, this allows the client to weigh in and the team to get clarity.
2. Large tickets are taken as bolts to be worked out in groups. If clarity is needed, then the team mobs to answer questions, or, the client is looped in to provide clarity.
3. We only demo every 2 weeks, like a traditional sprint. This allows the team to work through bolts quickly without being blocked waiting for client approval.
4.  

This isn’t a perfect method, and I’m sure we don’t get the higher end of 60% improvement. So, if you want to attempt a similar methodology, there are some things you may want to consider:

• What zoom level your tickets are at:With AI DLC you are able to take whole epics at once.
• Differences in skill level: Seniors will like the higher level of tickets and the speed at what they can develop. More junior members may get intimidated by the large chunk of work and will need more moving.
• Multiple Grooming sessions: With AI DLC, tickets will go fast, and because of this you may be tempted to spend more time up front grooming tickets. This is a trap. Groom enough for everyone to have 1-2 bolts, then break. Meet again if needed to get more started. It’ll allow you to put discoveries and context gained into the next units of work.

AI Development is changing at extreme rates, but having a framework that is cross-platform and cross-model is the best strategy to speed-up adoption, and value.